Abstraction via Separable Components: An Empirical Study of Absolute and Relative Accuracy in Processor Performance Modeling

With each new chip generation, high-end microprocessors become more complex and therefore more difficult to design, analyze, model and simulate. Developing and debugging cycle-accurate simulators for modern microprocessors is increasingly time-consuming. Likewise, using these simulators to explore a design space thoroughly can take months or more of simulation time. It is therefore increasingly important that CPU designers employ abstraction methodologies that allow them to make good microarchitectural design decisions without relying solely on many runs of cycle-accurate simulators. In this paper, we examine one popular modeling abstraction in detail. The separable components methodology separates program performance into an idealized base CPI (cycles per instruction) figure and other additive stall figures corresponding to key components of program performance. While this method is widely used by designers as a back-of-the-envelope trick, prior research has not thoroughly examined its accuracy characteristics. This paper studies the abstraction's accuracy in detail. Our primary finding is that separable components has excellent relative accuracy and is therefore highly effective at choosing appropriate design points from a design space, even in cases where its absolute accuracy is less impressive. Since the method provides upper and lower bounds on performance, it is also useful for early-state reasoning about a design's sensitivity to different parameters and design choices.